MLP Model for Emotion Recognition using Acoustic Features

Emotion recognition is an important topic of research lately. There are already a few methods that can predict emotion but, in this paper, we im to make a unique model that is not only lightweight but also fast and accurate. We are currently focussing on predicting 4 emotions like angry, sad, neutral and happy through frequencyanalysis. To achieve this, we first needto extract7 features including 195 sub-features from an audio file. The features like MFCC, Root mean square, Spectral contrast, Tonnetz, Zero-crossing rate, Mel spectrogram frequency and Chroma combined are useful in determining emotion from the audio frequencies accurately. We have trained the model using RAVDESS and TESS which are both open source databases for audio emotions. We made an intelligent python program that can analyze the frequency pattern of each emotion using MLP classifier and predict the emotion with an accuracy of 83%.Moreover, we can predict emotion either from a file or from the microphone instantly.